Workspace components and vulnerability metadata
Components require a controller and execution host with the current component protocol. An older or unsupported host reports unavailable. Saved configuration does not establish an installed scanner, a completed change or a vulnerability assessment.
The Components view belongs to devcontainer Templates, prepared Images and existing Containers. It uses the shared terminal/web/desktop interface. Optional managed tools can be selected or removed. Required runtime, selected agent and Git-provider dependencies, MCP tooling and management connections stay protected. Changing a workspace selection does not uninstall software on the host or the computer running the client.
Optional selection covers Tailscale, less, ripgrep, zip, unzip and the five managed context tools. Other installed packages are metadata, not an arbitrary uninstall interface.
Each row distinguishes declared configuration, observed installation and an unavailable observation. Installed and latest available versions are separate. A Feature version is not automatically the version of software it installs. Official release notes are opened through a curated link; this view does not ask a model to interpret changelogs.
Save starts the existing durable tooling operation. Image preparation and application to existing containers have independent progress and results. Each selected installation or deselected managed removal has its own progress and duration. An unknown step count uses an activity indicator. Closing the view does not cancel the execution-host job. Follow the existing operation/history view to distinguish queued work, verified installation, updated guides and native session notification acknowledgement. A notice does not establish tool usage.
Removing optional Tailscale
Tailscale is excluded from new workspace images unless selected. Existing installations require an exact managed target, supported ownership evidence and dependency checks. A familiar binary path alone is insufficient. Standalone Feature installations require byte-matched official artifacts; package-managed installations require the exact package/version and a successful dependency check. A byte-matched official legacy startup helper can be replaced with a fixed no-op while its entrypoint and persistent state remain intact. Unknown helpers are retained.
An active daemon or route, unverified startup dependency, changed executable, unknown ownership, shared package dependency or stale container identity blocks removal. The row explains the blocker; it does not report success after merely saving the checkbox. Rebuilding a prepared image is separate from adopting it for a running session. Active native sessions are preserved.
Host/local Tailscale, management SSH/tunnels and persistent tailnet state are outside the operation. No automatic package autoremove, state purge or arbitrary directory cleanup is performed.
Understanding vulnerability counts
Counts apply to an exact package version and recorded scanner/database coverage. The detail retains severity, advisory ID/aliases and source, available fixed versions, matching confidence, scanner version, database build time and observed time. Repeated dependency paths or equivalent advisory aliases are deduplicated for the same package/version. Suppressed findings remain distinct.
Unknown, unavailable, unsupported, partial and stale states are explicit. A missing severity is counted as unknown. No findings is not a security guarantee. The latest version gets advisory counts only when exact-version authoritative evidence exists; a release label or version string alone is insufficient.
Inventory and advisory matching run locally through pinned Syft/Grype binaries. The advisory adapter is a preview. It targets observed Debian 12 package metadata and managed npm packages on Linux amd64. Other coverage remains explicit. The scanner needs the exact cached public Node base from the managed image lock, real host-proc/cgroup resource evidence and a fresh bounded host-clock observation. It never substitutes a private project image as the public-network scanner runtime.
The public database update happens before package metadata enters the worker. All worker networks are disconnected before matching. The database uses a bounded ephemeral tmpfs; scanner binaries and owner-scoped inventory/report digests use bounded retained caches. Unknown private packages are not sent to external advisory APIs. Raw source paths, credentials and arbitrary SBOM metadata are not sent to public APIs or included in the normal console response. Context-tool deactivation uses the existing managed-link adapter; retained version caches follow its retention policy and are not deleted from active provider processes.
Server refresh and history
The controller schedules vulnerability metadata refresh twice a day without an open TUI or browser. Jobs are bounded, owner-scoped and resource-admitted. Each report records the exact inventory digest, scanner version and advisory database identity. Only a report matching the current inventory can supply current-version counts. A changed inventory remains unknown until its own matching completes. A failed or uncertain command is not duplicated merely because the console refreshes.
History and advisory details load bounded pages. Recorded temporary artifacts are cleaned under a disk/age policy; live work, current database pins, unrelated files, Docker images and volumes are preserved. Exhausted budget or missing scanner support is reported as unavailable.
Optional interpretation
If deterministic mapping leaves an unresolved identity, an available model chosen in Control Room settings can make one bounded interpretation per recipe/evidence digest. The current adapter requires an owned ready Claude account, an explicit model from its current catalog and a host with evidence-only task support. The isolated profile has only controlled evidence/result tools and no shell or filesystem execution. It receives only sanitized structured declarations and returns identity suggestions. It cannot claim an installed version, create a vulnerability result, authorize removal or execute a shell command. Deterministic verification is still required.
API and automation
The typed source routes are GET /v1/components/targets,
PUT /v1/components/selection, POST /v1/components/refresh,
POST /v1/components/scan, GET /v1/components/vulnerabilities and
GET /v1/components/history. GET/PUT /v1/components/policy configure owner
refresh limits; PUT /v1/components/team-policy remains canonical-operator-only.
Targets are registered template/image/container identities, not local paths.
The server derives ownership and enforces team policy and revision checks.
The CLI helper supports components list, apply, refresh,
scan, vulnerabilities and history. Session MCP derives its target from the
authenticated session. configure_template_components requires the existing
manual danger approval protocol; read methods do not grant operator authority.
Actual native discovery remains client-specific: the broker tool list alone
does not prove that a running native client has registered a newly added name.